2.Related Datasets
2.1 RGB-D Datasets
Texture-class objects
benchmark: Model based training, detection and pose estimation of texture-less 3D objects in heavily cluttered scenes. In ACCV, 2012.
15 texture-less objects represented by a color 3D mesh model
Each object is associated with a test sequence consisting of ~1200 RGB-D images, each of which includes exactly one
instance of the object.
objects have discriminative color, shape and/or size
Additional ground truth poses for all modeled objects: Learning 6D object pose estimation using 3D object coordinates. In ECCV, 2014
Latent-class hough forests for 3D object detection and pose estimation
present a dataset with 2 texture-less and 4 textured object
Each object a color 3D mesh model with a test sequence of over 700 RGB-D images.
several object instances with no to moderate occlusion, and with 2D and 3D clutter.
Recovering 6D object pose and predicting next-best-view in the crowd. In CVPR, 2016.
183 test images of 2 textured objects， appear in multiple instances in a challenging bin-picking scenario with heavy occlusion.
color 3D mesh models of another 6 textured objects and 170 test images depicting the objects placed on a kitchen table.
The Challenge and Willow datasets
Multimodal blending for high-accuracy instance recognition. In IROS, 2013
a set of 35 textured household objects
Each object 37 RGB-D from different views, a color point cloud obtained by merging the training images.
respectively contain 176 and 353 test RGB-D images of several objects in single instances placed on top of a turntable.
The Willow datasets also features distractor objects and object occlusion.
the TUW dataset
17 textured and texture-less objects appearing in 224 test RGB-D images.
The Rutgers dataset
color 3D mesh models for 24 mostly textured objects from the Amazon Picking Challenge 2015
captured in more than 10K test RGB-D images with various amounts of occlusion.
A global hypotheses verification method for 3D object recognition. In ECCV, 2012.
3D mesh models without color information of 35 household objects that are both textured and texture-less
50 test RGB-D images of table-top scenes with multiple objects in single instances, with no clutter and various levels of occlusion.
The BigBIRD dataset
125 mostly textured objects
For each object, the dataset provides 600 RGB-D point clouds, 600 high-resolution RGB images, and a color 3D mesh model
A large-scale hierarchical multi-view RGB-D object dataset.
with 300 common household objects captured on a turntable from three elevations.
250K segmented RGB-D images and 22 annotated video sequences with a few hundred RGB-D frames in each.
Ground truth is provided only in the form of approximate rotation angles for training images and in the form of 3D point labeling for test images.
A new benchmark for pose estimation with ground truth from virtual reality. Production Engineering,2014.
synthesized RGB-D images from simulated object manipulation scenarios involving 4 texture-less objects from the Cranfield assembly benchmark
2.2 Depth-only and RGB-only Datasets
Depth
Three-dimensional model-based object recognition and segmentation in cluttered scenes. TPAMI, 2006.
3D mesh models of 5 objects and 50 test depth images acquired with an industrial range scanner.
The test scenes contain only the modeled objects that occlude each other
Variable dimensional local shape descriptors for object recognition in range data. In ICCV, 2007.
The Desk3D dataset
Robust instance recognition in presence of occlusion and clutter. In ECCV, 2014.
3D mesh models for 6 objects which are captured in over 850 test depth images with occlusion, clutter and similarly looking distractor objects.
RGB image
Parsing IKEA Objects: Fine Pose Estimation. In ICCV, 2013.
objects being aligned with their exactly matched 3D model
A novel representation of parts for accurate 3D object detection and tracking in monocular images. In ICCV,2015.
3D CAD models and annotated RGB sequences with 3 highly occluded and texture-less objects.
Fast 6D pose estimation for texture-less objects from a single RGB image. In ICRA, 2016.
RGB sequences of 6 texture-less objects that are each imaged in isolation against a clean background and without occlusion.
2.3. Datasets for Similar Problems
3. The T-LESS Dataset
- a larger number of industry-relevant objects
- training images captured under controlled conditions
- test images with large viewpoint changes, objects in multiple instances affected by clutter and occlusion; including test cases that are challenging even for the state-of-the-art methods,
- images captured with a synchronized and calibrated triplet of sensors
- accurate ground truth 6D poses for all modeled objects
- two types of 3D models for each object.
image acquisition, camera calibration, depth correction, 3D object model generation and the ground truth pose annotation.
3.1. Acquisition Setup
3.2. Calibration of Sensors
3.3. Training and Test Images
3.4. Depth Correction
3.5. 3D Object Models
3.6. Ground Truth Poses
a dense 3D model of the scene was first reconstructed with the system of Steinbrücker
The CAD object models were then manually aligned to the scene model.
rendered into several selected high-resolution scene images from Canon
The final poses were distributed to all test images with the aid of the known camera-to-turntable coordinate transformations.